dongdonghe1998 has open-sourced Iterlay, which offers an infinite canvas integrating reference images, image generation, editing, version management, and asset organization. It supports CLI and Skill interfaces for invocation by Agents such as Codex and Claude Code; the application runs locally under the MIT license, requiring users to configure model services themselves. The GitHub repository address is https://github.com/Davied-H/iterlay.
Latent Space released Issue #1 of AINews, summarizing AI developments from October 15–16, 2026: Meta launched Muse Spark Open (816B total parameters/42B active MoE, 1M context, Apache-derived license) and Reflection Beam 1.0 (501B/23B MoE, Apache 2.0); NVIDIA introduced Nemotron-3.5-Super-VL (120B/12B) and Bolt (74B/9B); StepFun released Step-5-Preview (600B/27B, non-commercial license); China plans to implement export controls on models with over 2T parameters; Qwen reached 141.3M downloads on Hugging Face, approximately 6 times that of other labs; The Reddit community has divergent views on Muse Spark Open regarding safety tuning, inference support, and deployment costs.
CalledKingsley developed and open-sourced the Chrome extension Key to Magpie, which supports one-click import of API keys into the magpie desktop application from the token management pages of new-api (/token) or sub2api (/keys), automatically filling in name, id, endpoint, key, model list, catalog, and official website address; it requires loading via developer mode, and the project is hosted on GitHub.
KimataSun has open-sourced the personal navigation page project Flaris Navigation, based on Cloudflare Workers + KV. It supports zero-cost deployment, glassmorphism UI, Open-Meteo weather, lunar calendar with dynamic greetings, multi-level category backend management, and bilingual support for Chinese and English. An online Demo and GitHub repository are provided. Deployment requires binding the NAV_DB KV Namespace and setting the ADMIN_PASSWORD environment variable.
Biblio Times is an independently developed app for iPhone, iPad, and Mac that locally formats users' subscribed RSS/Atom feeds in the style of a morning newspaper, featuring front pages, columns, serif headlines, halftone images, and other print characteristics; it requires no account and involves no server relay, keeping all data entirely on-device; it supports iOS 18, iPadOS 18, and macOS 15, with a one-time purchase covering all three platforms at a launch price of $14.99 (until the end of October 2026), including the first three issues free.
Anthropic announced that starting October 29, 2026, Claude Haiku 4.5 will be removed from the model selector in Free, Pro, and Max plans, covering regular chats on web, desktop, and mobile as well as Claude Cowork and Design features; Haiku 4.5 remains available in Team and Enterprise plans and via API; existing user conversations are not deleted, but subsequent responses will be generated by Haiku 5.5; Haiku 5.5 was released on October 7, supports configurable thinking levels, has a context window of 1000000 tokens, maximum output of 128K tokens, outperforms Haiku 4.5 in coding and agentic tasks, and reduces average operating costs by approximately 75%.
The author manually distilled an "identity package" containing identity, rules, progress, and priorities, paired with four questions that can be judged as correct or incorrect. Tests were conducted on Claude and Grok separately to compare the effects of providing versus not providing the package; results showed that after providing the package, both models answered all four questions correctly and demonstrated consistent priority reasoning, verifying the feasibility of "identity travels with the person." The package was not automatically generated by a product, the experiment was completed via a relay channel, and the related protocol receipt-v1 has been open-sourced.
A developer released siltop v1.2.1, a terminal system monitor designed specifically for Apple Silicon (M1–M5). It supports real-time display of CPU/GPU/ANE/DRAM power consumption, per-cluster core (E/P/Super) clocks, temperatures, and fan speeds. All data comes from native macOS APIs (SMC, IOReport, Mach, etc.) without requiring sudo privileges, with an overhead of approximately 1.3% on a single core. Features include JSON/NDJSON/Prometheus/CSV output, tmux/SketchyBar integration, multi-Mac remote views, session recording/playback, and continuous performance stress testing. It can be installed via Homebrew, precompiled binaries, or Cargo, and runs on the macOS Apple Silicon platform.
itsVic developed and released the open-source desktop tool BiliFlow, supporting Windows x64 and Apple Silicon Mac installation packages (including built-in FFmpeg and recording engine); Linux can run from source or use the CLI. Features include batch downloading and filtering of Bilibili multi-part/collections/series/favorites, local archiving of video/audio/subtitles/danmaku/covers/metadata, resumable download tasks, live stream monitoring and automatic recording, segmentation and reconnection after disconnection, MP4 history export, and a unified task center. The project uses the MIT License.
The Beijing Chaoyang District People's Court ruled in the first instance that an insurance company illegally terminated the labor contract of Ms. Wu, a legal affairs employee. The court found that her uploading of organized and processed audit data tables to a third-party AI platform to assist with work did not constitute a "serious violation of rules and regulations": there was no basis for determining subjective intent by the company, no prior ban on AI usage, and no evidence of actual harm. The judgment systematically clarified for the first time the legal distinction between "theoretical risk" and "actual damage" in AI office scenarios, emphasizing that companies must establish rules beforehand rather than pursue accountability after the fact.
A developer released the open-source accounting PWA application Zenance, which supports pure frontend operation, stores data 100% locally in IndexedDB, and implements AES-GCM end-to-end encrypted synchronization via the Dropbox API. It features interpersonal accounts receivable/payable management, multi-currency support, budget tracking, JSON backup, and PIN code locking. The tech stack is React 19 + TypeScript + Vite + Tailwind CSS v4 + Zustand + Dexie.js, licensed under AGPL-3.0.
The Vector Institute in Canada and Radboud University in the Netherlands published a paper on arXiv titled "SpireBench: Evaluating Long-Term Strategic Capabilities of Agent Models with Slay the Spire," proposing the SpireBench benchmark, which uses the game's highest difficulty, A20, to systematically test 20 large models.
seventuning77 developed the free web tool groktexttospeech.com based on xAI's open Grok TTS interface. It supports selecting from 28 built-in voices, pasting up to 1000 characters of text to generate 24 kHz MP3s, and is compatible with tone tags such as [pause][laugh][sigh][breath] and <whisper><slow><emphasis>. Each voice has an independent page for easy comparison, accompanied by an API cost estimator calculated at $15 per million characters. No registration is required, text is not stored, and there are frequency limits of 20 calls per hour and 10000 characters per day; it is declared as an unofficial project.
On October 12, a document titled "Victim Statement and Evidence Compilation" publicly accused former ByteDance intern Tian Keyu of deliberately sabotaging model training tasks for Goku, LlamaGen, and others between March and July 2024. The alleged actions included tampering with PyTorch optimizer parameters (such as forcing maximize=True), gradient clipping logic, and DistributedSampler random seed generation. These operations were highly covert, causing training anomalies to manifest as poor performance rather than errors, and often involved dynamically adjusting sabotage methods during troubleshooting meetings based on colleagues' progress. A court's first-instance ruling found him in breach of contract and ordered compensation of 500,000 yuan; the case is currently in the second instance. After verifying the code against PyTorch documentation, Silicon Star confirmed that if the described behaviors are true, they constitute serious deliberate sabotage of the model training environment.
Ulanqab relies on advantages including an average annual temperature of 4.3°C, low-cost green electricity from the Western Inner Mongolia grid (approx. 0.32 yuan/kWh), and low latency of 5 milliseconds to Beijing. As of August 2026, it has signed 109 computing projects, deployed 172,000 PFLOPS of computing power, and contracted over 7,000,000 racks, attracting investments exceeding 500,000,000,000 yuan from companies such as Huawei, Alibaba, ByteDance, Kuaishou, and DeepSeek. Locally, skills training classes and targeted vocational college programs for electromechanical and HVAC talent have helped young people like Chen Gang transition from service industries to data center shift leader roles; potato farming also utilizes satellites and sensors for data-driven management.
woai110120130 released aiopsterm (Apache-2.0), an open-source AI operations terminal supporting unified SSH / bastion host / K8s / database access, with AI workflows bound to the current session, zero installation on target machines, and bidirectional MCP integration; it provides split-screen view, file management, PostgreSQL/MySQL/Redis clients, a K8s panel, and an AI session inbox; the macOS signed installer is notarized, and anonymous telemetry can be disabled.
Citigroup has prohibited third-party AI agents from logging into customer accounts or executing transactions; Bank of America restricts AI agent access to personal deposits and emphasizes that fund transfers out of accounts must undergo manual authorization; JPMorgan Chase also limits external AI agent integration citing risk concerns. These measures stem from banking industry fears that AI agents could automatically transfer idle funds to high-yield accounts, leading to a loss of $1.6 trillion in non-interest-bearing deposits, thereby impacting profits and financial stability.
Microsoft has launched Copilot for Legal, adding over 20 legal-specific AI skills for commercial customers of Microsoft 365 Copilot. It is integrated into the Word desktop version and the Cowork workspace, supporting tasks such as contract drafting and review, obligation tracking, and case timeline organization. It also connects to third-party legal tech services like Bloomberg Law, Thomson Reuters HighQ, and DocuSign via a plugin registry.
JetBrains announced that its local AI coding tool, Junie Local, has achieved Day 0 compatibility with the NVIDIA RTX Spark platform, with Windows support upgraded to stable release; generation speed reaches up to 55 tokens per second, supporting Qwen3.6-27B or JetBrains' fused model. Hardware requirements include an RTX Spark device or a GPU with Ampere architecture or newer and at least 24GB of VRAM, requiring Windows 11 and driver version 581+.